Model Cards

Transparency documentation for APHI AI models—scope, performance, equity, limitations, and governance.

Overview

Metrics below include placeholders pending final validation. Replace before external procurement or press usage.

Outbreak Anomaly Detector Pilot

Purpose: Early detection of anomalous respiratory syndrome clusters from syndromic surveillance feeds.

Performance & Fairness
MetricOverallHigh-SVILow-SVI
Sensitivity0.840.810.85
PPV0.720.700.73
Median Lead-Time (days)2.52.42.6
False Alerts / wk0.9

Placeholder – update with validated metrics + confidence intervals.

Limitations
  • Not tuned for ultra-rare syndromes.
  • Holiday demand shifts can inflate false positives.
  • Lead-time depends on upstream data latency.
Governance

Quarterly fairness audit; drift alert if sensitivity drops >5% vs prior quarter.

Version: 1.2 • Ref 8

Overdose Risk Stratifier Pilot

Purpose: Prioritize outreach for opioid overdose prevention (naloxone distribution, treatment engagement).

Performance
MetricOverall
AUROC0.79
Top 5% risk captures overdose %41%
Brier0.13

Illustrative only.

Fairness

Monitoring subgroup false negative parity; alert if gap >7 p.p. sustained over two retrains.

Limitations
  • Dependent on PDMP completeness.
  • Limited signal for first-time users.
  • Population-level prioritization; not clinical diagnosis.

Version: 0.9 • Ref 5

Vaccination Coverage Optimizer Early

Purpose: Recommend allocation of mobile clinics & outreach to maximize incremental coverage.

Simulation Outputs
ScenarioGain vs Baseline
Rural cluster+7.2%
Urban underserved+5.1%
Aggregate (5 sites)+6.0%

Simulation – awaiting field validation.

Limitations
  • Assumes stable outreach capacity.
  • No real-time misinformation shock modeling yet.
  • Equity weighting misconfiguration may reduce impact.

Version: 0.3 • Ref 6

Disclaimer

Metrics are illustrative placeholders until validated. Models are not for individual clinical decision-making.